]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/mm/madvise.c | |
3 | * | |
4 | * Copyright (C) 1999 Linus Torvalds | |
5 | * Copyright (C) 2002 Christoph Hellwig | |
6 | */ | |
7 | ||
8 | #include <linux/mman.h> | |
9 | #include <linux/pagemap.h> | |
10 | #include <linux/syscalls.h> | |
11 | #include <linux/mempolicy.h> | |
12 | #include <linux/page-isolation.h> | |
13 | #include <linux/hugetlb.h> | |
14 | #include <linux/falloc.h> | |
15 | #include <linux/sched.h> | |
16 | #include <linux/ksm.h> | |
17 | #include <linux/fs.h> | |
18 | #include <linux/file.h> | |
19 | #include <linux/blkdev.h> | |
20 | #include <linux/swap.h> | |
21 | #include <linux/swapops.h> | |
22 | ||
23 | /* | |
24 | * Any behaviour which results in changes to the vma->vm_flags needs to | |
25 | * take mmap_sem for writing. Others, which simply traverse vmas, need | |
26 | * to only take it for reading. | |
27 | */ | |
28 | static int madvise_need_mmap_write(int behavior) | |
29 | { | |
30 | switch (behavior) { | |
31 | case MADV_REMOVE: | |
32 | case MADV_WILLNEED: | |
33 | case MADV_DONTNEED: | |
34 | return 0; | |
35 | default: | |
36 | /* be safe, default to 1. list exceptions explicitly */ | |
37 | return 1; | |
38 | } | |
39 | } | |
40 | ||
41 | /* | |
42 | * We can potentially split a vm area into separate | |
43 | * areas, each area with its own behavior. | |
44 | */ | |
45 | static long madvise_behavior(struct vm_area_struct *vma, | |
46 | struct vm_area_struct **prev, | |
47 | unsigned long start, unsigned long end, int behavior) | |
48 | { | |
49 | struct mm_struct *mm = vma->vm_mm; | |
50 | int error = 0; | |
51 | pgoff_t pgoff; | |
52 | unsigned long new_flags = vma->vm_flags; | |
53 | ||
54 | switch (behavior) { | |
55 | case MADV_NORMAL: | |
56 | new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ; | |
57 | break; | |
58 | case MADV_SEQUENTIAL: | |
59 | new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ; | |
60 | break; | |
61 | case MADV_RANDOM: | |
62 | new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ; | |
63 | break; | |
64 | case MADV_DONTFORK: | |
65 | new_flags |= VM_DONTCOPY; | |
66 | break; | |
67 | case MADV_DOFORK: | |
68 | if (vma->vm_flags & VM_IO) { | |
69 | error = -EINVAL; | |
70 | goto out; | |
71 | } | |
72 | new_flags &= ~VM_DONTCOPY; | |
73 | break; | |
74 | case MADV_DONTDUMP: | |
75 | new_flags |= VM_DONTDUMP; | |
76 | break; | |
77 | case MADV_DODUMP: | |
78 | if (new_flags & VM_SPECIAL) { | |
79 | error = -EINVAL; | |
80 | goto out; | |
81 | } | |
82 | new_flags &= ~VM_DONTDUMP; | |
83 | break; | |
84 | case MADV_MERGEABLE: | |
85 | case MADV_UNMERGEABLE: | |
86 | error = ksm_madvise(vma, start, end, behavior, &new_flags); | |
87 | if (error) | |
88 | goto out; | |
89 | break; | |
90 | case MADV_HUGEPAGE: | |
91 | case MADV_NOHUGEPAGE: | |
92 | error = hugepage_madvise(vma, &new_flags, behavior); | |
93 | if (error) | |
94 | goto out; | |
95 | break; | |
96 | } | |
97 | ||
98 | if (new_flags == vma->vm_flags) { | |
99 | *prev = vma; | |
100 | goto out; | |
101 | } | |
102 | ||
103 | pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT); | |
104 | *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma, | |
105 | vma->vm_file, pgoff, vma_policy(vma)); | |
106 | if (*prev) { | |
107 | vma = *prev; | |
108 | goto success; | |
109 | } | |
110 | ||
111 | *prev = vma; | |
112 | ||
113 | if (start != vma->vm_start) { | |
114 | error = split_vma(mm, vma, start, 1); | |
115 | if (error) | |
116 | goto out; | |
117 | } | |
118 | ||
119 | if (end != vma->vm_end) { | |
120 | error = split_vma(mm, vma, end, 0); | |
121 | if (error) | |
122 | goto out; | |
123 | } | |
124 | ||
125 | success: | |
126 | /* | |
127 | * vm_flags is protected by the mmap_sem held in write mode. | |
128 | */ | |
129 | vma->vm_flags = new_flags; | |
130 | ||
131 | out: | |
132 | if (error == -ENOMEM) | |
133 | error = -EAGAIN; | |
134 | return error; | |
135 | } | |
136 | ||
137 | #ifdef CONFIG_SWAP | |
138 | static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start, | |
139 | unsigned long end, struct mm_walk *walk) | |
140 | { | |
141 | pte_t *orig_pte; | |
142 | struct vm_area_struct *vma = walk->private; | |
143 | unsigned long index; | |
144 | ||
145 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) | |
146 | return 0; | |
147 | ||
148 | for (index = start; index != end; index += PAGE_SIZE) { | |
149 | pte_t pte; | |
150 | swp_entry_t entry; | |
151 | struct page *page; | |
152 | spinlock_t *ptl; | |
153 | ||
154 | orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl); | |
155 | pte = *(orig_pte + ((index - start) / PAGE_SIZE)); | |
156 | pte_unmap_unlock(orig_pte, ptl); | |
157 | ||
158 | if (pte_present(pte) || pte_none(pte) || pte_file(pte)) | |
159 | continue; | |
160 | entry = pte_to_swp_entry(pte); | |
161 | if (unlikely(non_swap_entry(entry))) | |
162 | continue; | |
163 | ||
164 | page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE, | |
165 | vma, index); | |
166 | if (page) | |
167 | page_cache_release(page); | |
168 | } | |
169 | ||
170 | return 0; | |
171 | } | |
172 | ||
173 | static void force_swapin_readahead(struct vm_area_struct *vma, | |
174 | unsigned long start, unsigned long end) | |
175 | { | |
176 | struct mm_walk walk = { | |
177 | .mm = vma->vm_mm, | |
178 | .pmd_entry = swapin_walk_pmd_entry, | |
179 | .private = vma, | |
180 | }; | |
181 | ||
182 | walk_page_range(start, end, &walk); | |
183 | ||
184 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
185 | } | |
186 | ||
187 | static void force_shm_swapin_readahead(struct vm_area_struct *vma, | |
188 | unsigned long start, unsigned long end, | |
189 | struct address_space *mapping) | |
190 | { | |
191 | pgoff_t index; | |
192 | struct page *page; | |
193 | swp_entry_t swap; | |
194 | ||
195 | for (; start < end; start += PAGE_SIZE) { | |
196 | index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
197 | ||
198 | page = find_get_page(mapping, index); | |
199 | if (!radix_tree_exceptional_entry(page)) { | |
200 | if (page) | |
201 | page_cache_release(page); | |
202 | continue; | |
203 | } | |
204 | swap = radix_to_swp_entry(page); | |
205 | page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE, | |
206 | NULL, 0); | |
207 | if (page) | |
208 | page_cache_release(page); | |
209 | } | |
210 | ||
211 | lru_add_drain(); /* Push any new pages onto the LRU now */ | |
212 | } | |
213 | #endif /* CONFIG_SWAP */ | |
214 | ||
215 | /* | |
216 | * Schedule all required I/O operations. Do not wait for completion. | |
217 | */ | |
218 | static long madvise_willneed(struct vm_area_struct *vma, | |
219 | struct vm_area_struct **prev, | |
220 | unsigned long start, unsigned long end) | |
221 | { | |
222 | struct file *file = vma->vm_file; | |
223 | ||
224 | #ifdef CONFIG_SWAP | |
225 | if (!file || mapping_cap_swap_backed(file->f_mapping)) { | |
226 | *prev = vma; | |
227 | if (!file) | |
228 | force_swapin_readahead(vma, start, end); | |
229 | else | |
230 | force_shm_swapin_readahead(vma, start, end, | |
231 | file->f_mapping); | |
232 | return 0; | |
233 | } | |
234 | #endif | |
235 | ||
236 | if (!file) | |
237 | return -EBADF; | |
238 | ||
239 | if (file->f_mapping->a_ops->get_xip_mem) { | |
240 | /* no bad return value, but ignore advice */ | |
241 | return 0; | |
242 | } | |
243 | ||
244 | *prev = vma; | |
245 | start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
246 | if (end > vma->vm_end) | |
247 | end = vma->vm_end; | |
248 | end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
249 | ||
250 | force_page_cache_readahead(file->f_mapping, file, start, end - start); | |
251 | return 0; | |
252 | } | |
253 | ||
254 | /* | |
255 | * Application no longer needs these pages. If the pages are dirty, | |
256 | * it's OK to just throw them away. The app will be more careful about | |
257 | * data it wants to keep. Be sure to free swap resources too. The | |
258 | * zap_page_range call sets things up for shrink_active_list to actually free | |
259 | * these pages later if no one else has touched them in the meantime, | |
260 | * although we could add these pages to a global reuse list for | |
261 | * shrink_active_list to pick up before reclaiming other pages. | |
262 | * | |
263 | * NB: This interface discards data rather than pushes it out to swap, | |
264 | * as some implementations do. This has performance implications for | |
265 | * applications like large transactional databases which want to discard | |
266 | * pages in anonymous maps after committing to backing store the data | |
267 | * that was kept in them. There is no reason to write this data out to | |
268 | * the swap area if the application is discarding it. | |
269 | * | |
270 | * An interface that causes the system to free clean pages and flush | |
271 | * dirty pages is already available as msync(MS_INVALIDATE). | |
272 | */ | |
273 | static long madvise_dontneed(struct vm_area_struct *vma, | |
274 | struct vm_area_struct **prev, | |
275 | unsigned long start, unsigned long end) | |
276 | { | |
277 | *prev = vma; | |
278 | if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP)) | |
279 | return -EINVAL; | |
280 | ||
281 | if (unlikely(vma->vm_flags & VM_NONLINEAR)) { | |
282 | struct zap_details details = { | |
283 | .nonlinear_vma = vma, | |
284 | .last_index = ULONG_MAX, | |
285 | }; | |
286 | zap_page_range(vma, start, end - start, &details); | |
287 | } else | |
288 | zap_page_range(vma, start, end - start, NULL); | |
289 | return 0; | |
290 | } | |
291 | ||
292 | /* | |
293 | * Application wants to free up the pages and associated backing store. | |
294 | * This is effectively punching a hole into the middle of a file. | |
295 | * | |
296 | * NOTE: Currently, only shmfs/tmpfs is supported for this operation. | |
297 | * Other filesystems return -ENOSYS. | |
298 | */ | |
299 | static long madvise_remove(struct vm_area_struct *vma, | |
300 | struct vm_area_struct **prev, | |
301 | unsigned long start, unsigned long end) | |
302 | { | |
303 | loff_t offset; | |
304 | int error; | |
305 | struct file *f; | |
306 | ||
307 | *prev = NULL; /* tell sys_madvise we drop mmap_sem */ | |
308 | ||
309 | if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB)) | |
310 | return -EINVAL; | |
311 | ||
312 | f = vma->vm_file; | |
313 | ||
314 | if (!f || !f->f_mapping || !f->f_mapping->host) { | |
315 | return -EINVAL; | |
316 | } | |
317 | ||
318 | if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE)) | |
319 | return -EACCES; | |
320 | ||
321 | offset = (loff_t)(start - vma->vm_start) | |
322 | + ((loff_t)vma->vm_pgoff << PAGE_SHIFT); | |
323 | ||
324 | /* | |
325 | * Filesystem's fallocate may need to take i_mutex. We need to | |
326 | * explicitly grab a reference because the vma (and hence the | |
327 | * vma's reference to the file) can go away as soon as we drop | |
328 | * mmap_sem. | |
329 | */ | |
330 | get_file(f); | |
331 | up_read(¤t->mm->mmap_sem); | |
332 | error = do_fallocate(f, | |
333 | FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, | |
334 | offset, end - start); | |
335 | fput(f); | |
336 | down_read(¤t->mm->mmap_sem); | |
337 | return error; | |
338 | } | |
339 | ||
340 | #ifdef CONFIG_MEMORY_FAILURE | |
341 | /* | |
342 | * Error injection support for memory error handling. | |
343 | */ | |
344 | static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end) | |
345 | { | |
346 | struct page *p; | |
347 | if (!capable(CAP_SYS_ADMIN)) | |
348 | return -EPERM; | |
349 | for (; start < end; start += PAGE_SIZE << | |
350 | compound_order(compound_head(p))) { | |
351 | int ret; | |
352 | ||
353 | ret = get_user_pages_fast(start, 1, 0, &p); | |
354 | if (ret != 1) | |
355 | return ret; | |
356 | ||
357 | if (PageHWPoison(p)) { | |
358 | put_page(p); | |
359 | continue; | |
360 | } | |
361 | if (bhv == MADV_SOFT_OFFLINE) { | |
362 | pr_info("Soft offlining page %#lx at %#lx\n", | |
363 | page_to_pfn(p), start); | |
364 | ret = soft_offline_page(p, MF_COUNT_INCREASED); | |
365 | if (ret) | |
366 | return ret; | |
367 | continue; | |
368 | } | |
369 | pr_info("Injecting memory failure for page %#lx at %#lx\n", | |
370 | page_to_pfn(p), start); | |
371 | /* Ignore return value for now */ | |
372 | memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED); | |
373 | } | |
374 | return 0; | |
375 | } | |
376 | #endif | |
377 | ||
378 | static long | |
379 | madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev, | |
380 | unsigned long start, unsigned long end, int behavior) | |
381 | { | |
382 | switch (behavior) { | |
383 | case MADV_REMOVE: | |
384 | return madvise_remove(vma, prev, start, end); | |
385 | case MADV_WILLNEED: | |
386 | return madvise_willneed(vma, prev, start, end); | |
387 | case MADV_DONTNEED: | |
388 | return madvise_dontneed(vma, prev, start, end); | |
389 | default: | |
390 | return madvise_behavior(vma, prev, start, end, behavior); | |
391 | } | |
392 | } | |
393 | ||
394 | static int | |
395 | madvise_behavior_valid(int behavior) | |
396 | { | |
397 | switch (behavior) { | |
398 | case MADV_DOFORK: | |
399 | case MADV_DONTFORK: | |
400 | case MADV_NORMAL: | |
401 | case MADV_SEQUENTIAL: | |
402 | case MADV_RANDOM: | |
403 | case MADV_REMOVE: | |
404 | case MADV_WILLNEED: | |
405 | case MADV_DONTNEED: | |
406 | #ifdef CONFIG_KSM | |
407 | case MADV_MERGEABLE: | |
408 | case MADV_UNMERGEABLE: | |
409 | #endif | |
410 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
411 | case MADV_HUGEPAGE: | |
412 | case MADV_NOHUGEPAGE: | |
413 | #endif | |
414 | case MADV_DONTDUMP: | |
415 | case MADV_DODUMP: | |
416 | return 1; | |
417 | ||
418 | default: | |
419 | return 0; | |
420 | } | |
421 | } | |
422 | ||
423 | /* | |
424 | * The madvise(2) system call. | |
425 | * | |
426 | * Applications can use madvise() to advise the kernel how it should | |
427 | * handle paging I/O in this VM area. The idea is to help the kernel | |
428 | * use appropriate read-ahead and caching techniques. The information | |
429 | * provided is advisory only, and can be safely disregarded by the | |
430 | * kernel without affecting the correct operation of the application. | |
431 | * | |
432 | * behavior values: | |
433 | * MADV_NORMAL - the default behavior is to read clusters. This | |
434 | * results in some read-ahead and read-behind. | |
435 | * MADV_RANDOM - the system should read the minimum amount of data | |
436 | * on any access, since it is unlikely that the appli- | |
437 | * cation will need more than what it asks for. | |
438 | * MADV_SEQUENTIAL - pages in the given range will probably be accessed | |
439 | * once, so they can be aggressively read ahead, and | |
440 | * can be freed soon after they are accessed. | |
441 | * MADV_WILLNEED - the application is notifying the system to read | |
442 | * some pages ahead. | |
443 | * MADV_DONTNEED - the application is finished with the given range, | |
444 | * so the kernel can free resources associated with it. | |
445 | * MADV_REMOVE - the application wants to free up the given range of | |
446 | * pages and associated backing store. | |
447 | * MADV_DONTFORK - omit this area from child's address space when forking: | |
448 | * typically, to avoid COWing pages pinned by get_user_pages(). | |
449 | * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking. | |
450 | * MADV_MERGEABLE - the application recommends that KSM try to merge pages in | |
451 | * this area with pages of identical content from other such areas. | |
452 | * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others. | |
453 | * | |
454 | * return values: | |
455 | * zero - success | |
456 | * -EINVAL - start + len < 0, start is not page-aligned, | |
457 | * "behavior" is not a valid value, or application | |
458 | * is attempting to release locked or shared pages. | |
459 | * -ENOMEM - addresses in the specified range are not currently | |
460 | * mapped, or are outside the AS of the process. | |
461 | * -EIO - an I/O error occurred while paging in data. | |
462 | * -EBADF - map exists, but area maps something that isn't a file. | |
463 | * -EAGAIN - a kernel resource was temporarily unavailable. | |
464 | */ | |
465 | SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior) | |
466 | { | |
467 | unsigned long end, tmp; | |
468 | struct vm_area_struct *vma, *prev; | |
469 | int unmapped_error = 0; | |
470 | int error = -EINVAL; | |
471 | int write; | |
472 | size_t len; | |
473 | struct blk_plug plug; | |
474 | ||
475 | #ifdef CONFIG_MEMORY_FAILURE | |
476 | if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE) | |
477 | return madvise_hwpoison(behavior, start, start+len_in); | |
478 | #endif | |
479 | if (!madvise_behavior_valid(behavior)) | |
480 | return error; | |
481 | ||
482 | if (start & ~PAGE_MASK) | |
483 | return error; | |
484 | len = (len_in + ~PAGE_MASK) & PAGE_MASK; | |
485 | ||
486 | /* Check to see whether len was rounded up from small -ve to zero */ | |
487 | if (len_in && !len) | |
488 | return error; | |
489 | ||
490 | end = start + len; | |
491 | if (end < start) | |
492 | return error; | |
493 | ||
494 | error = 0; | |
495 | if (end == start) | |
496 | return error; | |
497 | ||
498 | write = madvise_need_mmap_write(behavior); | |
499 | if (write) | |
500 | down_write(¤t->mm->mmap_sem); | |
501 | else | |
502 | down_read(¤t->mm->mmap_sem); | |
503 | ||
504 | /* | |
505 | * If the interval [start,end) covers some unmapped address | |
506 | * ranges, just ignore them, but return -ENOMEM at the end. | |
507 | * - different from the way of handling in mlock etc. | |
508 | */ | |
509 | vma = find_vma_prev(current->mm, start, &prev); | |
510 | if (vma && start > vma->vm_start) | |
511 | prev = vma; | |
512 | ||
513 | blk_start_plug(&plug); | |
514 | for (;;) { | |
515 | /* Still start < end. */ | |
516 | error = -ENOMEM; | |
517 | if (!vma) | |
518 | goto out; | |
519 | ||
520 | /* Here start < (end|vma->vm_end). */ | |
521 | if (start < vma->vm_start) { | |
522 | unmapped_error = -ENOMEM; | |
523 | start = vma->vm_start; | |
524 | if (start >= end) | |
525 | goto out; | |
526 | } | |
527 | ||
528 | /* Here vma->vm_start <= start < (end|vma->vm_end) */ | |
529 | tmp = vma->vm_end; | |
530 | if (end < tmp) | |
531 | tmp = end; | |
532 | ||
533 | /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */ | |
534 | error = madvise_vma(vma, &prev, start, tmp, behavior); | |
535 | if (error) | |
536 | goto out; | |
537 | start = tmp; | |
538 | if (prev && start < prev->vm_end) | |
539 | start = prev->vm_end; | |
540 | error = unmapped_error; | |
541 | if (start >= end) | |
542 | goto out; | |
543 | if (prev) | |
544 | vma = prev->vm_next; | |
545 | else /* madvise_remove dropped mmap_sem */ | |
546 | vma = find_vma(current->mm, start); | |
547 | } | |
548 | out: | |
549 | blk_finish_plug(&plug); | |
550 | if (write) | |
551 | up_write(¤t->mm->mmap_sem); | |
552 | else | |
553 | up_read(¤t->mm->mmap_sem); | |
554 | ||
555 | return error; | |
556 | } |